ollama

What this local model job needs

People who type ollama usually want a free local model runner they can install once and keep between experiments. The job is pulling a model, chatting with the CLI, and serving an API on a PC you control, not a browser demo and not a metered cloud seat.

The upstream open-source project behind that search ships a desktop tray app plus a CLI. Herd Latch maps the doors: the Windows Setup executable on GitHub Releases tags, the verified winget id, and the Homebrew desktop cask when those channels are allowed.

Related: releases, GitHub, download safely, Windows install.

Why this desktop install path

A dedicated install stays on the machine after Setup. You control the update door. Model libraries, Modelfiles, and the local API live under your account instead of a temporary browser tab that forgets your lab layout.

Winget keeps fleet images honest when the catalog is allowed. The Setup executable keeps labs honest when package managers are blocked. Brew keeps Mac desks on the same project without hunting mirrors. The Linux install script and amd64 archive keep Linux desks on the same tag as Windows and macOS.

Proprietary GUI runners still win when a team already standardised on another brand with a polished model browser. Cloud chat seats do not replace an offline-capable local runner. Choose this MIT path when you want named Releases assets and CLI habits that start with a short prove command.

Install by operating system

Windows, macOS, and Linux all ship stranger-safe doors on the same stable tags. Prefer the package manager when it is allowed; fall back to the named Releases asset when catalogs are blocked.

Run the Windows winget line when App Installer is present:

winget install -e --id Ollama.Ollama

On macOS, prefer the verified desktop cask:

brew install --cask ollama-app

On Linux, the official install script is the common stranger-safe path:

curl -fsSL https://ollama.com/install.sh | sh

After install, prove the CLI with a small model:

ollama run llama3.2

Longer Windows pages: install on Windows. Winget-only walkthrough: winget install. Installer walkthrough: Windows installer. Cluster pages: releases, GitHub, download safely, how to install winget.

How this client compares

People comparing this runner to other local tools pick by licence, platforms, and how much UI they want on day one. Numbers below are product facts, not rankings.

LM Studio still wins for people who want a dense GUI catalog first. GPT4All wins when a team already ships those packs. llama.cpp wins when you are assembling a custom stack. This MIT runner wins when you want Releases assets and a short path to a CLI prove chat.

Head-to-head: vs LM Studio. Leaving a GUI rival: switch from LM Studio. More comparisons live under guides.

First-hour steps after install

These eight steps match a clean Windows desk. macOS and Linux follow the same prove-one-chat habit after their own installer doors.

1. Open the official release door. Open GitHub Releases for the upstream repository on a current non-prerelease tag and confirm you see the Windows Setup executable, not a renamed portal setup.
2. Prefer the Windows Setup from the current tag. Take the Windows Setup executable from the live non-prerelease tag that ships it. Keep one filename written on the machine ticket. The asset is large; expect a long download.
3. Or run the verified winget id. When App Installer is present, run the verified winget install command for this project. That id is YAML-verified in microsoft/winget-pkgs and maps to the Setup executable.
4. Finish the Windows Setup wizard. Approve UAC when Windows prompts. Complete Setup so the tray app and CLI land on PATH.
5. Confirm the tray or CLI responds. Launch the app and confirm the tray icon appears, or run the version command in a new terminal before you pull models.
6. Run one model to prove the install. Run a small model chat with the CLI and send one short prompt. That proves download, serve, and chat before you trust larger models.
7. Note the local library path. Models land under the local library on disk. Confirm free space before large pulls, especially on lab images with small system volumes.
8. Keep one update door. Return to the same GitHub Releases Setup file, winget upgrade, or brew cask upgrade. Skip adware wrappers that only rank for local AI searches.

First-run detail: first run. After install checklist: installed.

Safe download habits for ollama

Official doors are GitHub Releases for the upstream repository, the verified winget id, and the verified Homebrew cask. Those publish named assets. Advertising mirrors often rename the Setup and bundle extra offers.

Match the filename you expect before you run anything. The primary Windows stranger-safe file is the Setup executable on stable tags. A portable zip may also ship; take it only when your runbook documents a portable need.

Deep dive: download safely. Repo map: GitHub overview. Asset chooser: releases.

Models, Modelfiles, and everyday CLI

After install, the everyday habit is pulling a tag and starting a chat. List local models, then remove stale tags when disk pressure rises. Custom Modelfiles let you pin system prompts without leaving the MIT runner.

Examples you can run on your machine after install:

ollama run llama3.2

API clients can talk to the local server after the service is up. Keep that loop on loopback until your network policy says otherwise. Guide: first run.

Lab and travel notes

Travel laptops benefit from a single documented installer door written beside the folder that holds the library. Lab images drift when someone installs a second mirror during a blocked SmartScreen prompt. Prefer returning to the same Releases URL, winget id, or brew cask instead of inventing a new download mid-flight.

School machines should receive the installer on hardware they control, plus a shared note for the update door. Shared desks should keep a short prove-one-chat checklist before calling the image ready.

That habit catches vanished install folders and adware shortcuts that reappear from old bookmark exports. Users who already live in a GUI rival can keep that path for browsing. Treat the Windows Setup as the dedicated path for teammates who need a named Releases asset with CLI control on day one.

Support rhythms for shared rooms

Helpdesks move faster when every ticket names the installer door and the Windows build together. Vague reports that only say chats feel wrong waste a day. Ask for the filename or winget id, the approximate install date, and whether a proxy changed during the complaint.

• Keep a spare USB with the current Setup executable for rooms without winget.
• Use a quiet prove-one-chat step before demoing a huge model pull.
• Inventory model allow-lists separately from the desktop binary.

A forgotten PATH refresh looks like a broken client. Tag hardware notes with the matching library path so substitutes do not guess under pressure. When imaging season arrives, freeze the golden machine only after a reboot test. Logon scripts and antivirus first-scan delays hide problems that a same-session demo misses. The boring reboot test saves weekend emergency calls.

Writing internal runbooks

Internal runbooks should quote the winget id, the brew cask, and the Releases URL without adding marketing adjectives. Staff skim under stress. A short checklist with five boxes outperforms a three-page essay that nobody finishes.

Include a rollback line: where the previous Setup lives, how to uninstall, and who owns the software library hash list. Rollback plans fail when the only copy sat on a retired laptop.

Translate jargon for volunteers. Say “use the official GitHub Setup file” instead of assuming everyone knows which mirror is safe. Clarity beats cleverness during the first week of a new semester.

How do I install the Windows build?

Prefer the Windows Setup from the current non-prerelease tag on the upstream GitHub Releases, or run the verified winget install command when App Installer is present. Both doors map to the same MIT build for a local model runner. Avoid renamed setups from advertising download portals that only rank for the head term.

Where should I download the installer?

Use GitHub Releases for the upstream repository, the verified winget package, or the verified Homebrew cask. Those doors publish named assets on stable tags. Softonic-style portals often wrap the same search intent with extra installers you did not ask for.

Is the desktop app free?

The runner is free under MIT. You can install and use it without buying a seat. Model weights you pull follow each model’s licence. Optional cloud services from other vendors are separate from this install path.

Can I switch from LM Studio into this client?

Yes. Install from GitHub Releases, winget, or brew, then pull models with the CLI. There is no automatic LM Studio library import. Prove one small model first. Guide: switch from LM Studio.

Does the desktop app work offline?

After you install the app and download each model you need, chats can run without the public internet. You need a network to pull new models or to fetch a newer Setup executable from Releases.

Profile and model habits that age well

Keep a short allow-list of model tags for shared rooms. Large experimental pulls fill disks and confuse the next class. Document which tags are blessed for demos. Remove stale tags during maintenance windows.

Custom Modelfiles belong in version control when a course depends on a fixed system prompt. Treat the Modelfile like any other lab config: review it when the base model tag changes.

API scripts should target loopback by default. Opening the server beyond the machine needs an explicit network decision, not a silent demo flag.

Imaging season without drama

Golden images should pin a non-prerelease tag that still ships the Windows Setup executable. Record the tag on the image build sheet. After sysprep or clone, open a terminal and prove a small CLI chat before you seal the image.

Winget-based images should confirm the package id resolves in the catalog used by that ring. Brew-based Mac images should pin the desktop cask. Linux images should document the install script versus an offline archive.

Writing clearer tickets

Tickets that say “AI broken” waste time. Ask for the OS, the door used (winget, Setup executable, brew, install script), the tag, and the exact CLI error. Attach version command output when possible.

If the failure is a pull timeout, note proxy settings. If the failure is disk full, note the library path free space. If the failure is PATH, ask whether the terminal was opened before Setup finished.

Classroom logistics for ollama labs

Scheduling ollama installs across a lab bank works better when one person stages the Windows Setup overnight and copies it to trusted media before students arrive. Partial downloads create support noise that looks like security failures but is usually truncated bits.

• Document which ring owns the blessed tag for ollama images.
• Volunteers need a five-box checklist: door, tag, prove chat, free space, update door.
• Travel carts should carry the Setup file and a SmartScreen filename note.

Shared accounts should avoid experimental multi-gigabyte tags on the golden profile. Scratch accounts exist for curiosity. Maintenance windows delete strays so the next ollama class starts with a known library.

When two local AI apps share a disk, write which one is blessed for the course. Students click whichever shortcut is larger. Clarity beats clever branding on the desktop.

Print the Releases URL and the winget show command on the ollama lab sheet. Phones photograph well. Whiteboard URLs get erased overnight.

More install and setup guides

Browse the guide hub for OS installs, comparisons, and switch-from pages. Every guide links back to this pillar.

• Install by operating system

  Step-by-step installs with the exact command or file for each system.

  • Install on Windows
  • Install on macOS
  • Install on Linux
  • Install on Ubuntu
  • winget install
  • Windows installer
  • Homebrew cask

• Comparisons

  ollama against other local model runners people ask about.

  • vs GPT4All
  • vs Jan

• Switch and import

  Move from another local AI app into ollama without rival-first titles.

  • ollama: switch from GPT4All

• Setup and models

  First run, CLI ollama run examples, safety, and updates.

  • First run
  • Is it safe?
  • Update and uninstall